AutoGraphForge: Towards Automated Graph Theory Discovery

TEXT START: We report on our ongoing project to develop a computational pipeline, AutoGraphForge, for an automated graph-theoretic conjecturing-refuting-formalizing-proving system.

The Dissection

This is an industrialization prototype for mathematical labor. It converts conjecture generation, counterexample hunting, formal statement construction, and proof search into a repeatable machine pipeline. The headline number—6,522 surviving conjectures—is throughput, not validated discovery. “Survived” means only that candidates escaped the listed datasets, active searches, novelty filter, and current proving process. It does not mean they are true, important, general, or useful.

The Lean kernel is a genuine correctness barrier, but only for the formal statement and proof supplied to it. It does not guarantee that the formalization captures the intended mathematical claim, that the custom invariant preamble is adequate, or that the resulting theorem has intellectual value. The hand-proved examples demonstrate feasibility, not autonomous mathematical competence.

The Core Fallacy

The text treats pipeline completion and formal verification as proxies for mathematical discovery. They are not. Finite refutation datasets cannot establish universal truth, and a novelty filter built from 559 relations cannot define novelty outside its encoded closure.

Its deeper omission is economic: it measures whether the system runs, not how much human mathematical labor it displaces or whether it scales cheaply enough to dominate that labor. Under the Discontinuity Thesis, that omission is decisive. The machine does not need to replace all mathematicians at once. It only needs to make enough routine conjecturing, testing, and proof construction cheaper and faster than human execution. This project is one mechanism for that erosion.

Hidden Assumptions

• The selected graph families and random models are representative of the relevant conjecture space.
• Failure to find a counterexample is treated as meaningful evidence rather than an artifact of coverage.
• The encoded invariant language can express the discoveries worth making.
• The novelty filter correctly distinguishes new mathematics from consequences it cannot encode.
• Automated translation into Lean preserves the intended semantics.
• Kernel verification is being mistaken for verification of the research question itself.
• HPC expense, dataset construction, and invariant computation will remain economically tolerable at larger scale.
• The few relations proved by hand are evidence of a general pipeline rather than carefully selected successes.
• Human researchers remain embedded in the architecture, invariant design, dataset choice, and interpretation, but that dependency is not counted as a continuing labor bottleneck.
• “Currently running” and “initial sanity checks” are treated as evidence of an achieved result, although the complete end-to-end outcome is not yet reported.

Social Function

Primary classification: partial truth. Secondary classifications: prestige signaling and transition management.

The project genuinely demonstrates that a bounded portion of mathematical work can be mechanized, iterated, and kernel-checked. Its prestige function is to turn that mechanization into a credible research frontier. Its transition-management function is subtler: it normalizes the conversion of mathematicians from discoverers and provers into designers, curators, validators, and owners of computational infrastructure.

The resulting message is not “mathematics is already dead.” It is more dangerous: substantial pieces of mathematical participation are becoming machine-runnable before institutions have acknowledged what that implies.

The Verdict

AutoGraphForge is not proof of a machine mathematician. It is proof of an emerging factory for mathematical candidate production and proof labor. Its current limitations are real, but they are lag defenses—finite datasets, narrow domains, compute cost, and human-designed scaffolding—not reversals of the underlying direction. If cognitive automation achieves durable cost and performance superiority, systems like this will not preserve the human wage role in mathematics; they will help hollow it out.